Startup raises $650 million to create AI that learns and improves its own code

Startup Recoursive Superintelligence Inc., composed of former Google, Meta and OpenAI members, has just announced an investment of $650 million. The stated objective is ambitious: to create artificial intelligence models that can analyze and improve their own source code without human intervention. The round, one of the largest in the sector in 2026, signals that the market bets heavily on the idea of self-sufficient AI — and that the race for superintelligence is getting closer and closer to the real world.

How does self-performing AI work?

The core concept of the project is the creation of models with "recursion" capability — that is, they can perform self-assessment loops and generate better versions of themselves. Instead of relying on human engineers to adjust parameters or fix bugs, the AI performs an internal process of continuous training. The Recursive Superintelligence team states that their first prototypes have already managed to reduce the optimization time of neural structures in controlled tests by 40%, according to data published in the ad.

This is different from traditional fine-tuning techniques: here, artificial intelligence itself defines the rules of evolution, detects performance bottlenecks and rewrites parts of the code autonomously. The approach resembles the concept of AutoML, but taken to the extreme — a system that redesigns itself.

Why the investment in self-performing AI relevant to Brazil?

The volume of resources and origin of the founders place Recursive Superintelligence at the center of a trend that affects the entire technology ecosystem. For Brazilian developers and companies, the news reinforces the importance of following open models and tools that allow auditability. If the AI starts to modify its own code, transparency becomes a strategic asset — especially in regulated sectors such as health, finance and law.

In addition, startup can inspire similar movements in emerging innovation hubs. The algorithmic autonomy race is not exclusive to Silicon Valley: startups like Mira Murati Startup Presents AI with Real-time Communication(/noticias/startup-de-mira-murati-apresenta-ia-com-comunicacao-em-tempo-real) show that Brazil can also benefit from solutions that reduce dependence on huge engineering teams.

What are the risks and challenges of self-performing AI?

The promise of an AI that improves itself raises questions about control and security. If the system starts to change its behavior unpredictablely, how do we ensure that it remains aligned with human goals? The Recursive Superintelligence team itself admits that external supervision mechanisms will be essential — and that investment will also serve to develop internal "security brakes".

Another computational challenge is: training recursive models consumes processing power far above average. Startup has already signed partnerships with dedicated data centers to scale infrastructure. The scientific community follows the case carefully, because the success or failure of these methods can define the next generation of AI architectures.

Frequently Asked Questions about self-performing AI

What does "IA that improves the code itself" mean?

It is a type of artificial intelligence that, during training or execution, can identify weaknesses in its own programming and generate new, more efficient versions without needing a human engineer to rewrite the lines.

Is this model available for use?

Not yet. Recursive Superintelligence is in the process of research and development, with internal prototypes. The forecast is that the first commercial applications will emerge in 2027, focused on optimizing large language models.

How long before we see fully autonomous AI?

Experts differ. Some believe the advance can be exponential, with recursive systems reaching superintelligence in less than a decade. Others warn that theoretical and security barriers can delay this scenario indefinitely.

Read the original news in SiliconAngle for more technical details.